1. Viral Video Style: A Closer Look at Viral Videos on YouTube Lu Jiang, Yajie Miao, Yi Yang, Zhenzhong Lan, Alexander G. Hauptmann School of Computer Science, Carnegie Mellon University

2. Outline  Introduction  CMU Viral Video Dataset  Statistical Characteristics  Peak Day Prediction  Conclusions

3. Outline  Introduction  CMU Viral Video Dataset  Statistical Characteristics  Peak Day Prediction  Conclusions

4. What is a viral video? A viral video is a video that becomes popular through the process of (most often) Internet sharing through social media. Gangnam Style

5. What is a viral video? A viral video is a video that becomes popular through the process of (most often) Internet sharing through social media. Gangnam Style Charlie bit my finger Missing pilot MH370

6. Social Validity Viral videos have been having a profound impact on many aspects of society. Politics: – Pro-Obama video “Yes we can” went viral (10 million views) in 2008 US presidential election [Broxton 2013]. – Obama Style and Mitt Romney Style went viral (30 million views in the month of Election Day), and peaked on Election Day.

7. Social Validity Viral videos have been having a profound impact on many aspects of society. Politics: – Pro-Obama video “Yes we can” went viral (10 million views) in 2008 US presidential election [Broxton 2013]. – We found that Obama Style and Mitt Romney Style went viral (30 million views in the month of Election Day), and peaked on Election Day.

8. Social Validity

9. Social Validity cont. Viral videos have been having a profound impact on many aspects of society. Financial marketing: – Old Spice’s campaign went viral and improved the brand’s popularity among young customers[West et al 2011]. – Psy’s commercial deals has amounted to 4.6 million dollars from Gangnam Style.

10. Social Validity cont. Viral videos have been having a profound impact on many aspects of society. Financial marketing: – Old Spice’s campaign went viral and improved the brand’s popularity among young customers[West et al 2011]. – Psy’s commercial deals has amounted to 4.6 million dollars from Gangnam Style.

11. Motivation Existing studies are conducted on: – Small set (tens of videos)  biased observations? – Large-scale Google set  confidential! A relatively large and public dataset on viral videos would be conducive. Solution: CMU Viral Video Dataset. Beware the content are hilarious!

12. Outline  Introduction  CMU Viral Video Dataset  Statistical Characteristics  Peak Day Prediction  Conclusions

13. CMU Viral Video Dataset By far the largest public viral video dataset. Time’s list was the largest dataset (50 videos). Videos are manually selected by experts – Editors from Time Magazine, YouTube and the viral video review episodes. Statistics: 446 viral videos, 294 quality 19,260 background videos. 10+ million subscribers!

14. CMU Viral Video Dataset Cont. For a video, it includes: – Thumbnail – Video and user metadata – Insight data: historical views, likes, dislikes, etc. – Social data: #Inlinks, daily tweeter metions (pending) – Near duplicate videos (automatic detection + manual inspection).

15. CMU Viral Video Dataset Cont. For a video, it includes: – Thumbnail – Video and user metadata – Insight data: historical views, likes, dislikes, etc. – Social data: #Inlinks, daily tweeter mentions (pending) – Near duplicate videos (automatic detection + manual inspection).

16. CMU Viral Video Dataset Cont. For a video, it includes: – Thumbnail – Video and user metadata – Insight data: historical views, likes, dislikes, etc. – Social data: #Inlinks, daily tweeter mentions (pending) – Near duplicate videos (automatic detection + manual inspection).

17. CMU Viral Video Dataset Cont. For a video, it includes: – Thumbnail – Video and user metadata – Insight data: historical views, likes, dislikes, etc. – Social data: #Inlinks, daily tweeter mentions (pending) – Near duplicate videos (automatic detection + manual inspection).

18. Outline  Introduction  CMU Viral Video Dataset  Statistical Characteristics  Peak Day Prediction  Conclusions

19. Statistical Characteristics Observations agree with existing studies including the study on Google’s dataset – Short title, Short duration. Less biased.

20. Observation I The days-to-peak and the lifespan of viral videos decrease over time. Days-to-peak Lifespan

21. Observation II The popularity of the uploader is a more substantial factor that affects the popularity than the upload time. Upload time is believed to be the most important factor in background videos[Borghol et al. 2012], coined as First-mover advantage. An example: official music videos

22. Outline  Introduction  CMU Viral Video Dataset  Statistical Characteristics  Peak Day Prediction  Conclusions

23. Peak Day Prediction Forecast when a video can get its peak view based on its historical view pattern. ???

24. Peak Day Prediction Forecast when a video can get its peak view based on its historical daily view pattern. ???

25. Peak Day Prediction Forecast the date a video can get its peak view based on its daily view pattern. ???

26. Peak Day Prediction Forecast when a video can get its peak view based on its historical daily view pattern. True peak date

27. Peak Day Prediction Forecast the date a video can get its peak view based on its daily view pattern. This application is significant in supporting and driving the design of various services: – Advertising agencies: determine timing and estimate cost – YouTube: recommendation – Companies/Politicians: respond viral campaigns

28. HMM Model Model daily views using HMM model: Two types of states: – hibernating  less views – active  more views h

29. HMM Model Model daily views using HMM model: Two types of states: – hibernating  less views – active  more views h h

30. HMM Model Model daily views using HMM model: Two types of states: – hibernating  less views – active  more views h h a1a1

31. HMM Model Model daily views using HMM model: Two types of states: – hibernating  less views – active  more views h h a2a2 a1a1

32. HMM Model Model daily views using HMM model: Two types of states: – hibernating  less views – active  more views h h a2a2 a1a1 a3a3 h h

33. HMM Model Model daily views using HMM model: Two types of states: – hibernating  less views – active  more views Novel modifications: Incorporate metadata in the prediction. Other work only use the pure view count [Pinto et al. 2013]. Smooth transition probability by a Gaussian prior.

34. Experimental Results Considering metadata in peak day prediction is instrumental.

35. Result cont. Early warning system for viral videos. Detect viral videos and forecast their peak dates.

36. Outline  Introduction  CMU Viral Video Dataset  Statistical Characteristics  Peak Day Prediction  Conclusions

37. Conclusions A few messages to take away from this talk: – CMU Viral Video Dataset is by far the largest open dataset for viral videos study. – This paper discovers several interesting characteristics about viral videos. – This paper proposes a novel method to forecast the peak day for viral videos. The preliminary results look promising.

38. References T. West. Going viral: Factors that lead videos to become internet phenomena. Elon Journal of Undergraduate Research, pages 76–84, 2011. T. Broxton, Y. Interian, J. Vaver, and M. Wattenhofer. Catching a viral video. Journal of Intelligent Information Systems, 40(2):241–259, 2013. Y. Borghol, S. Ardon, N. Carlsson, D. Eager, and A. Mahanti. The untold story of the clones: content-agnostic factors that impact youtube video popularity. In SIGKDD, pages 1186–1194, 2012. H. Pinto, J. M. Almeida, and M. A. Gon¸calves. Using early view patterns to predict the popularity of youtube videos. In WSDM, pages 365–374, 2013.